New framework tests when AI citation data is actually reliable

Rank correlation trajectories and confidence intervals are not the most glamorous subject in B2B marketing. The arXiv paper "From Stochastic to Stable," published in July 2025, makes them load-bearing.

The paper's argument is blunt: most AI visibility measurement is conducted without any principled test of whether the data collected is sufficient to support the conclusions drawn. Practitioners run a set of prompts, tally which domains a generative search engine cites, rank them, and proceed. The framework introduced here calls that procedure into question. It proposes two sequential tests. The first, rank stability, asks whether the rank-correlation trajectory across successive prompt batches has flattened into a genuine plateau or is still drifting. The second, structural sufficiency, asks whether the gaps between domains' citation shares are large enough relative to the uncertainty of those estimates to support any inference at all.

That second criterion is the more demanding one. A ranking can stabilise and still be useless for decisions: if the confidence intervals around each domain's citation share overlap substantially, the apparent order between domain A and domain B is noise dressed as signal.

The measurement gap that practitioners have been ignoring

The practical consequence is uncomfortable. A substantial share of the "brand visibility" reports circulating in marketing and communications teams are probably built on rankings that would fail the structural sufficiency test. Collection budgets vary enormously. A study that runs 200 prompts through Perplexity and a study that runs 2,000 prompts through ChatGPT will produce rankings of very different reliability, but both are typically presented with identical confidence. The framework formalises the difference.

For financial services firms, multilateral institutions, and large industrial groups, this matters acutely. These organisations increasingly commission or consume AI visibility audits to understand how models like ChatGPT or Google's AI Overviews position them relative to competitors and peer institutions. A bank that sees itself ranked third behind two rivals in an LLM citation audit might restructure its content strategy in response. If that ranking is based on insufficient data, the restructuring is responding to measurement error.

The framework also exposes a platform-specific problem. Generative search engines differ in how stochastically they produce citations: one model may cite the same domain reliably across near-identical prompts while another varies considerably. A fixed prompt-count that achieves structural sufficiency on one platform may fall well short on another. Studies that aggregate across platforms without accounting for this are doubly unreliable.

Who benefits from this, and who is exposed

Tooling vendors in the AI visibility space have a material interest in the outcome here. Firms selling citation-tracking products to enterprise clients are, implicitly, asserting that their sample sizes are adequate. The framework provides a public standard against which those claims can now be tested. Vendors whose methodologies survive the stability and sufficiency tests gain a credible differentiator. Those whose products rest on thin prompt samples face an awkward conversation.

For in-house teams, the more immediate implication is about procurement and interpretation. Before acting on any AI visibility ranking, two questions now have defensible technical answers: has the rank order stopped moving with additional data, and are the citation-share gaps between ranked domains wide enough to mean anything? The paper derives both criteria from the statistical regularities of the underlying citation distributions, which means the tests are tractable, not just theoretically appealing.

There is a harder implication beneath this. The entire practice of measuring brand visibility in LLM outputs is younger than most enterprise budget cycles. Standards for what constitutes adequate evidence have not caught up with commercial demand for the data. The sequential convergence framework is, among other things, a reminder that moving fast on AI visibility strategy while ignoring measurement quality is not boldness. It is a different kind of risk.

Organisations that set the methodological bar now, insisting on sufficiency tests before acting on citation rankings, will make better decisions with the same data. Those that don't will keep optimising for positions that may not exist.